Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S965706AbdGTVWo (ORCPT ); Thu, 20 Jul 2017 17:22:44 -0400 Received: from gate.crashing.org ([63.228.1.57]:56255 "EHLO gate.crashing.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S934535AbdGTVWn (ORCPT ); Thu, 20 Jul 2017 17:22:43 -0400 Message-ID: <1500585713.10674.10.camel@kernel.crashing.org> Subject: Re: [PATCH] drivers: cpuidle: Disable preemption before get_lppaca function call in pseries_idle_probe function From: Benjamin Herrenschmidt To: Victor Aoqui , rjw@rjwysocki.net, daniel.lezcano@linaro.org, paulus@samba.org, mpe@ellerman.id.au, linux-pm@vger.kernel.org, linuxppc-dev@lists.ozlabs.org, linux-kernel@vger.kernel.org Cc: ltc-interlock@lists.linux.ibm.com, victora@br.ibm.com, mauricfo@linux.vnet.ibm.com Date: Fri, 21 Jul 2017 07:21:53 +1000 In-Reply-To: <20170720175735.3957-1-victora@linux.vnet.ibm.com> References: <20170720175735.3957-1-victora@linux.vnet.ibm.com> Content-Type: text/plain; charset="UTF-8" X-Mailer: Evolution 3.24.3 (3.24.3-1.fc26) Mime-Version: 1.0 Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 1314 Lines: 32 On Thu, 2017-07-20 at 14:57 -0300, Victor Aoqui wrote: > When CONFIG_PREEMPT=y, the following warning shows up: > > BUG: using smp_processor_id() in preemptible [00000000] code: swapper/0/1 > caller is pseries_processor_idle_init+0x58/0x21c > > This warning shows up because preemption cannot occur when using > get_paca(), otherwise the paca_struct it points to may be the wrong one > just after. > > For this reason, preemption needs to be disabled before > lppaca_shared_proc(get_lppaca()). Also chekc the generated assembly. We had all sort of interesting issues where gcc would copy the paca pointer or the lppaca pointer to a GPR *outside* of the preempt disabled section... In that specific case it's not a big deal but overall, I am not comfortable with PREEMPT on powerpc until we do something a bit more drastic... I would like to remove all such direct accesses to paca, instead have a "new" get_paca() written in asm that does the preempt disable then returns the PACA in a GPR (not directly use r13, hide that from gcc), and which is paired with a put_paca(). The few places where we want to directly access r13 should be hand written in asm too to hide r13 from gcc, for accessing the irq_happened in the fast path of local_irq_enable/disable/... we should do the same with lock tokens. Ben.